Assessing the cross-region knowledge transfer capability of selected deep learning building vectorization methods in...

Area of Science:

• Earth Observation
• Computer Vision
• Deep Learning

Background:

• Building vectorization from Earth Observation imagery is crucial for urban mapping.
• State-of-the-art deep learning models are increasingly used for this task.
• Assessing cross-region knowledge transfer is vital for deploying models in data-scarce areas.

Purpose of the Study:

• To evaluate the cross-region knowledge transfer capabilities of deep learning-based building vectorization methods.
• To determine the impact of network architecture and training dataset choice on transferability.
• To identify optimal model-dataset combinations for processing Earth Observation data in regions with limited training data.

Main Methods:

• Investigated Convolutional Neural Network (CNN), Transformer, and Graph Neural Network (GNN) architectures, alongside general-purpose models.
• Utilized diverse datasets (SpaceNet AOI 2, SpaceNet AOI 5, CrowdAI, WHU-Mix) for training.
• Tested models on a custom dataset from Dar es Salaam, Tanzania, to assess performance in a low-data region.

Main Results:

• Cross-region transferability is primarily influenced by segmentation and vertex selection methods, not just training performance.
• High-quality, diverse training datasets from different regions yield superior results compared to lower-quality, local datasets.
• SAMPolyBuild trained on WHU-Mix showed the best overall performance, while Pix2Poly excelled in individual accuracy metrics.

Conclusions:

• Model architecture and training data significantly impact the success of cross-region knowledge transfer in building vectorization.
• Prioritizing diverse, high-quality training data is essential for robust performance in under-resourced geographic areas.
• Specific architectures like SAMPolyBuild and Pix2Poly demonstrate potential for effective transfer learning in remote sensing applications.